JP3413236B2 - Character recognition method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method of forming a synthetic discriminant function having a high discriminating ability in recognizing a character such as character information written on an X-ray film by using the synthetic discriminant function. The present invention relates to a character recognition method using the composite identification function and a preprocessing method for character recognition.

[0002]

2. Description of the Related Art Currently, most medical images are X-ray films, and digitization of X-ray films is automatically performed by a digitizer. Input is currently being done manually. In a large hospital, the number of X-ray films that must be processed in one day has risen to several hundreds, and under the current circumstances, a great deal of effort has to be put into inputting text information on the film. The development of a system that automatically recognizes is desired.

[0003]

Various methods have been attempted for character recognition, but it is very difficult to achieve a recognition rate of 100% in character recognition, and the well-known Bayes decision rule, etc. There is a problem that it takes a long time to recognize when trying to increase the recognition rate using.

[0004]

The present invention has been made in view of such conventional problems, and an object of the present invention is to improve the accuracy of character recognition and enable recognition in a short time .

[0006]

Therefore, according to the present invention,
The character recognition method has two classes for the types of characters to be recognized.
, And each combination of the classification is shown by the following formula.
The synthetic knowledge formed to maximize the Fischer ratio
A group of different functions is stored in the computer and the computer
The data is processed by the computer and the
When the combination of a predetermined classification with a large shear ratio
Select the discriminant function, and then sort by the composite discriminant function
Of the two classes that can be classified,
Classify into two classes and set a predetermined class as above.
Select the composite discriminant function corresponding to the matching, and select
Characteristic of recognizing all kinds of characters by repeating .

[Equation 2]

[0007]

[Equation 2]

[0008]

In that case, the group of composite discriminant functions first forms a first composite discriminant function for discriminating one character from the rest of the characters, and then one of the remaining characters. A second composite discriminant function for discriminating the remaining characters excluding the character may be formed, and the same operation may be repeated thereafter to recognize all the characters.

Then, a correlation calculation with a character to be recognized is performed using each composite discriminant function formed as described above , and a class having a larger correlation value is selected to change from a large class to a small class. They are classified and recognized as specific characters. Further, the probability of causing an error in the identification result may be obtained and stored based on the correlation value obtained by the correlation calculation with each synthetic identification function used until the character is recognized.

[0011]

When it is determined that a character is erroneously recognized, the composite discriminant function having the highest probability of error in the stored discrimination result is detected and the discrimination result is changed. The characters may be recognized again by re-identifying. In this case, the character combination obtained by recognizing each character may be collated with the recorded data of the correct character combination for confirmation, so that the character recognized by mistake may be identified. . Also, the present invention
Character information is recorded as a character recognition method for
The step of detecting a character area that is detected, and the detected character
Extracting character information from the area, and
Character information and multiple composite discriminant functions corresponding to the character
The calculation is performed and the correlation value based on each calculation result is output.
Force step, and based on the output result of the correlation value,
Select one from the plurality of composite discriminant functions
Identifying steps and database of identified characters
Performing a data collation in, by the comparison
If they do not match, based on the output of the composite discriminant function,
To convert the specified character to a different character.
And-flops, may have a.

[0012]

[0013]

[0014]

When a plurality of different characters are identified by the correlation calculation with the composite identification function, even if the correlation with the same character with different font is high, but the decorrelation with other characters is not high, it is not always necessary. It cannot be said that it has a good identification function. In particular, if the correlation with the same character is increased as the number of fonts increases, the non-correlation with other characters decreases. Therefore, it can be said that the composite discriminant function having high discriminative ability is determined by a balance between high correlation with respect to the same character and high non-correlation with respect to different characters.

In the Fischer ratio used as the evaluation criterion of the composite discriminant function in the present invention, the value of the numerator represents the degree of difference in the correlation value between the two classes to be discriminated, and the value of the denominator is the degree of dispersion of the correlation value in each class. Represents the total of. In other words, the higher the correlation with respect to the same character, the smaller the degree of dispersion of the correlation value, the smaller the value of the denominator, and the higher the decorrelation of different characters, the greater the degree of difference in the correlation value. The larger the value, the larger the Fischer ratio. Therefore, when the condition that the correlation with the same character is high and the decorrelation with the different character is high, the Fischer ratio becomes the maximum and the identification function also becomes the maximum.

From the above points, by forming a composite discriminant function using the Fischer ratio as an evaluation criterion, a composite discriminant function having a high discriminating function can be obtained. Further, if only two characters are to be identified, one synthetic identification function may be simply formed so that the Fischer ratio is maximized. It is necessary to consider the number of functions and the time required for identification.

For example, from the large class to the small class, for example, the above 10 kinds of characters are classified into 2 classes and the classified characters are further classified into 2 classes according to the identification procedure such as a so-called tree structure. There is a method of classifying and classifying these classifications by a composite classification function. As a method of forming a composite discriminant function in that case, a composite discriminant function is formed so that the Fischer ratio is maximized for all possible combinations of classifications, and the classification having the maximum Fischer ratio among these is formed. If a combination of is selected, it is possible to obtain a composite discriminant function in which the classification / discrimination performance is improved as much as possible in each of large and small classes.

Therefore, by performing character recognition by using the composite discriminant function formed by the above-mentioned method, high accuracy
Moreover, character recognition can be performed in a short time. In the above case, if each synthetic discriminant function is formed so as to discriminate one character from the rest of the characters, the probability of recognition with a smaller number of times of discrimination is improved. Further, by storing the probability that an error will occur in the identification result based on the correlation value with each composite identification function, when it is found that an error occurs later, based on the data of the misidentification probability. The identification result of the composite identification function, which has a high probability of incorrect identification, can be changed and identification can be performed again. It is more efficient to detect such a composite discriminant function having the maximum false discrimination probability and automatically perform discrimination again.

On the other hand, apart from the tree structure, there is also a method of forming a composite discriminant function corresponding to each character. That is, each composite discriminant function discriminates a predetermined character from other characters and has a high correlation with the predetermined character and a high non-correlation with other characters. There is a need. Also in this case, by forming the Fischer ratio as the evaluation reference, it is possible to form a group of composite discriminant functions each having a discriminant function enhanced as much as possible.

When character recognition is performed using this group of synthetic discriminant functions, the character to be recognized is subjected to a correlation operation with respect to each synthetic discriminant function and recognized by selecting the one having the maximum correlation value. Is done. Further, when it is found that the recognition has an error by this method, the probability of correct correction can be increased by selecting and recognizing the one having the next highest correlation value.

Further, in order to ensure character recognition for medical purposes, it is sufficient to compare the recognized character combination with a separately recorded correct character combination, and the recognized character is recorded. If it is not present in the data, it will prove to be erroneous and thus can be re-identified as described above .

[0022]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 shows a configuration of a character information automatic recognition system that recognizes character information written on an X-ray film by using a composite discriminant function and performs preprocessing by the method according to the present invention prior to the character information recognition. Show.

In the figure, when the X-ray film F irradiated with X-rays and having the character information written in the peripheral portion is set in the digitizer 1, the digitizer 1 digitizes the X-ray image and the character information data to obtain the image data. Is input to the processing computer 2 which performs character recognition according to the present invention.
The process of automatic recognition of character information by the processing computer 2 is roughly divided into preprocessing and character recognition / collation as shown in the figure.

In the preprocessing, first, the area on the input X-ray film F in which the character information is copied is specified, the inclination of the character is detected and corrected, and the film insertion direction is detected to detect the image. After correcting to the regular position, cut out the characters. In the character recognition / collation, a character is extracted by using a composite identification function (hereinafter referred to as SDF) to form character information, and an error collection code (ECC) such as a checksum attached to the character information is formed. Is used to check, if correct, and if there is a mistake, the letters are replaced, and then collated with the database of patient information stored in the radiation information system 3.

Next, the details of each process will be described. In the pre-processing, the characters necessary for recognizing the character information are cut out from the input X-ray film image data. There are several types of X-ray film depending on their size. For example, a 14-inches long and 14-inches long X-ray film called a large angle has a sampling interval of 0. 175 mm, digitized into 2000 × 2000 pixels with 10-bit gradation. The textual information includes the examination date, patient ID, patient name, date of birth, and sex. The pre-processing is roughly divided into four processes of specifying the character information area, correcting the inclination of the character, detecting the insertion direction of the film, and cutting the character, and the processing at each stage will be described.

The first step of the pre-processing is to specify the area on the X-ray film where the character information is written. FIG. 2 shows a list of places where the character information of the large-angle X-ray film used in the hospital is copied. There are nine candidate locations, but there are a total of four candidate locations because X-ray film must be digitized upside down or upside down as shown in FIG.
It reaches 0 places. There are other types of X-ray film, such as large four, six-section, and half-section, and the location where the character information is imprinted and the size are different depending on the type. Furthermore, the method of imprinting differs depending on the hospital, and detailed measures are required.

Therefore, in the present invention, vertical and horizontal lines as shown in FIG. 4 are added to the character information portion in order to easily detect the inclination of the character and judge the input state of the film. That is,
The vertical line is added to the left side of the character information, and the horizontal line is added to the lower side of the character information. The process of identifying the character information area will be described with reference to the flowchart shown in FIG.

The variance value of the image data is calculated according to the following equation for each candidate area registered in the memory (step 1).

[0029]

[Equation 3]

The area where the calculated variance value is maximum is detected as the character information area. The fact that the variance value of a portion where there is an edge such as a character and there is shading is large is used (step 2). Here, in consideration of being erroneously detected, the top three areas from the largest one are stored as candidate areas, the presence or absence of the components of the vertical and horizontal lines are detected at the same time, and if the components are not detected, 2nd and 3rd
Switch to the candidate area.

Next, in order to match the long side of the area with the horizontal direction, the image data is rotated so that the area having the maximum dispersion value is located on the upper side of the left corner (step 3). Here, in the case where the region is located at the corner and includes both sides that are perpendicular to each other, even if the region is located at the upper side of the left corner, the long sides of the region may coincide with each other in the vertical direction, and therefore it cannot be determined. In such a case, the long sides and the short sides may be distinguished by comparing the maximum values in the vertical and horizontal projection processing described later.

In the second step, the inclination of the character is detected and the inclination is corrected. In the current X-ray inspection, an X-ray inspection technician manually inserts a card, which is called a name card, into which the attribute information of the patient is written, into the machine, and optically burns it onto an X-ray film to record the patient information on the film. This is the format for recording. Because it is a method to insert the name card by hand,
Some characters are significantly inclined with respect to the film,
It may cause the character cutout failure.
Therefore, it is necessary to correct the inclination of the character.

The processing for detecting and correcting the inclination of the character will be described with reference to the flowchart shown in FIG. The differential operation represented by the following equation is performed on the data (density data) in the specified character information description area (step 11). f (i, j) = [ {f (i + 1, j) -f (i, j)} 2 + {f (i, j +1) -
f (i, j)} ² ] ^{1/2 The} n-valued (threshold processing) is performed on the differential image subjected to the differential operation (step 12).

A projection process is performed on the image data that has undergone such a process (step 13). The projection process is a process of obtaining the distribution of the total value of the data when the image data is projected in one direction as shown in FIG. Initially, the projection direction is set to the horizontal (horizontal) direction. Then, a predetermined tilt angle ± β with respect to the horizontal direction
Projection processing in the projection direction is performed while inclining each unit angle α to (for example, ± 4 °) (step 14 → step 15 → step 13). The inclination is performed by performing affine transformation on the image. In the affine transformation, the gradation value of the pixel (x, y) of the image after conversion is determined by the gradation value of the pixel (x ', y') of the image before conversion, which is determined by the following equation. .

[0035]

[Equation 4]

However, n = 1 to β / α After completion of the projection process, the angle θ having the maximum value of the total data value in the projection process in each projection direction is obtained (step 16). Here, when the line in the character arrangement direction coincides with the projection direction, the total value of the line data becomes maximum, so the projection direction having the maximum value coincides with the direction of the line in the character arrangement direction. Therefore, the angle θ formed by the projection direction and the horizontal direction can be obtained as the inclination angle of the character. Here, it goes without saying that the smaller the unit angle α, the higher the detection accuracy, but the interpolation can also increase the accuracy.

The tilt is corrected by rotating the image by the affine transformation by the tilt angle θ of the character thus obtained (step 17). The third step is to judge the insertion direction of the X-ray film. As described above, there are cases where the top and bottom and the front and back are opposite to the normal insertion direction of the X-ray film, so these are discriminated and corrected to the normal position.

The up / down and front / back correctness determination processing will be described with reference to the flowcharts shown in FIGS. After performing the differential operation (step 21), the n-value conversion (step 22), and the horizontal projection process (step 23) on the image for which the tilt correction is performed, it is determined whether the maximum value is above or below. Yes (step 24). In a simple manner, it is possible to proceed to step 24 and make the determination by using the projection image having the maximum value in the projection processing performed while rotating the image in steps 13 to 15.

Then, when it is determined that the maximum value is on the lower side, the process is finished as it is, but when it is determined that it is on the upper side, the vertical direction of the image is reversed (step 25).
That is, since the horizontal line is written on the lower side of the character information, if the horizontal line detected as the maximum value is located on the upper side, it is inverted upside down to correct the vertical direction. be able to.

Next, projection processing in the vertical (vertical) direction is performed (step 26), and it is determined whether the maximum value is on the left or right (step 27). Then, when it is determined that the image is on the left side, the process ends as it is, but when it is determined that the image is on the right side, the horizontal direction of the image is reversed (step 28). That is, since the vertical line is written on the left side of the character information, when the vertical line detected as the maximum value is located on the right side, the horizontal direction should be reversed to correct the horizontal direction. Can (step
29).

By the above processing, the image is set in the normal direction without inclination. The fourth stage is the cutting of characters. That is, it is to detect an area in which each character is described when recognizing a character. The character cutting process will be described with reference to the flowchart shown in FIG.

The image corrected in the input direction as described above is subjected to the differential operation and the n-value conversion, and then the vertical and horizontal projection processes are performed (steps 31 and 32).
. However, since the projection processing in the vertical and horizontal directions is performed on the image after the tilt correction at the time of correcting the input direction without newly performing such projection processing, it is necessary to reset the data in the correct direction. May be.

By the vertical and horizontal projection processing, the character description portion has a large sum of the data amount and thus becomes a mountain shape, and the portion between the characters is a blank because the data value becomes small and becomes a valley. , The area surrounded by the vertical and horizontal grid axes passing through the valley is defined as one character area, the cut-out position of each character is determined (step 33), and the character is cut out based on the determined cut-out position ( Step 34).

After the preprocessing is performed as described above, the cut out characters are recognized and collated. The character recognition / collation processing will be described with reference to the flowchart shown in FIG. First, the result of reading the character information is output by performing a correlation calculation between the cut out character and the SDF (step 41). In short, the correlation calculation is to calculate the inner product of the input character pattern vector and the SDF vector, and the correlation value is the inner product value obtained by the correlation calculation. It is necessary to recognize 10 characters from 0 to 9 as the character information. In the present embodiment, 10 SDFs are combined using the 10 characters as a standard pattern, and the correlation value is different for each character pattern. +1 is set to +1 and other patterns are set to -1.

Then, a correlation operation is performed on one character and the above 10 SDFs, and the output values are compared to obtain the most +1.
The pattern is determined by examining the SDF that has output a correlation value close to (step 42). Figure 11 shows the schematic diagram. In the character information read by the SDF, the redundant part of the character information is checked, and if it is satisfied, then it is collated with the database (step 43). For example,
The checksum added to the character information is used for the redundancy of the character information. The checksum is for determining the number of the digit as a function of other numbers by making a digit of the character information redundant and has an error detecting ability. If there is a matching number in the database, the character information is recognized as having been correctly read. If there is not, correct the digit with the character information as a reading error.
After checking the checksum (step 44), the verification is performed again (step 45). The digit to be corrected and the character to be corrected are determined by comparing the correlation output values of SDF.
If there is no matching number even if it is corrected to some extent, it is considered unreadable (steps 46 and 47). In addition, the collation also collates with the collated character information for preventing erroneous recognition, and when there are two or more sets of corresponding numbers, both are output as unreadable.

To identify a character of a new font, the SDF added with a standard pattern corresponding to the font may be modified and additionally registered. Computer simulation was performed for each stage of the recognition process. First, characters were cut out from an actual X film as a test pattern, and the correct reading rate of character recognition was examined. The recognition method is as described above. The correct reading rate was 98.9%. A simulation of collation and correction of character information (ID number) was performed based on this correct reading rate. A database is created in which 1000 ID numbers satisfying the checksum are randomly created. The same number as each ID number was created by combining the above test patterns, and the IDs were identified by SDF, and then simulations of collation and correction were performed. Correction of characters was performed up to 3 digits of the 7-digit ID number. The results are shown in Table 1. The ID number that was made unreadable in the simulation was because there were four or more misreading characters, and it could not be corrected. From the result that the correct reading rate for each character is 98.9%, the probability that all 7 digits are correctly read is 92.5%. It was shown that this method can improve the recognition rate of ID numbers to 99.7%. From this result, it can be said that the above algorithm is effective.

[0047]

[Table 1]

Experiments were also carried out through a series of processes in which the X-ray film was digitized, the ID number was read, the ID number was checked against the database, and then recognized. This is to confirm the problems and changes to be clarified when the system proposed in this research is operated by using an actual film. The photographic film used in the experiment has the vertical and horizontal lines in the character information portion.
This was digitized without any restrictions on how to input the film, and recognition was performed. The results are shown in Table 2. As a result of testing 127 sheets of recognition, 124 sheets (97.6 sheets)
%) Met. The three causes of being unreadable are ID
The number of the last digit of the number was imprinted on the edge of the character area and the contrast was poor, so it could not be read correctly even after pre-processing. Considering that these X-ray films were created for this simulation and are in bad conditions such as imprinting a name card obliquely than usual, the recognition rate is considered to be further improved in actual use. .

[0049]

[Table 2]

Next, an embodiment of a method for forming an optimum SDF used for character recognition while evaluating it with a Fischer ratio will be described. First, the function of the SDF will be described. Assuming that the number of fonts is M, the pattern of M fonts for the same character is represented by an N-dimensional vector as the following equation.

[0051]

[Equation 5]

By thus determining the standard pattern and its output value, the SDF can be designed. The conditional expression (2) regarding the SDF expressed by the expression (3) means that the SDF is limited to the vector in the subspace defined by the M standard patterns of the N-dimensional space, and is derived from this.
The SDF of equation (3) is not a general solution of mathematical linear equation (1). In contrast, Z. Bahri et al.
In 1988, we derived a generalized SDF as a general solution of the equation of linear equation (1), and showed that various SDFs that had been derived so far can be expressed using the generalized SDF.

The conventional SDF (conventional SDF) and the generalized SDF are expressed as follows by the singular value decomposition method (SVD).

[0054]

[Equation 6]

Further, the deviation from the set correlation value due to noise that is not additive to the standard pattern and is not object-dependent
F and the generalized SDF are expressed by the SVD expression as follows.

[0056]

[Equation 7]

Since the generalized SDF is designed also by using the component orthogonal to the standard pattern, it has a feature that the degree of freedom in design is much larger than that of the conventional SDF. From the shift due to the noise, the variances of the conventional SDF and the generalized SDF are shown below.

[0058]

[Equation 8]

Here, the variance of the deviation from the set value of the conventional SDF depends only on the standard pattern, but the generalized SDF
The DF also depends on the component that is orthogonal to the standard pattern. Paying attention to this point, in the generalized SDF, if the component orthogonal to the standard pattern is determined so as to satisfy the following equation, a stable SDF can be designed even with general colored noise.

[0060]

[Equation 9]

Therefore, this β _i is determined using the Fischer ratio as an evaluation criterion, and an optimum SDF is formed. The Fischer ratio, which is a function of the linear discriminant function h, is expressed by the following equation.

[0062]

[Equation 10]

The linear discriminant function that maximizes the Fischer ratio is the projection axis of the multidimensional space that most emphasizes the separation of the two classes, and the Fischer ratio is maximized assuming that the pattern has a Gaussian distribution in the multidimensional space. SD that is more optimal than the conditions
F can be formed (see FIG. 12). Now, the Fischer ratio can be written as follows using the above h _g . If there are M ₁ and M ₂ standard patterns in classes 1 and ₂ , respectively, and each pattern of each class is represented as x ₁ ^(i), etc., and the average vector is represented by a bar, the denominator is ,

[0064]

[Equation 11]

The eigenvector having the maximum eigenvalue is h _{g obtained} by solving this eigenvalue problem. However, in the case of the Fischer ratio of 2 classes, the right side of the equation (6) from the defining equation (4) of D is

[0066]

[Equation 12]

We conclude that In this way, it was clarified that the SDF can be optimally designed by using the Fischer ratio as an evaluation criterion. A comparison between the generalized SDF formed using the Fischer ratio as an evaluation criterion and the conventional SDF will be shown below. The comparison was made for the identification problem of two numbers "0" and "1". As for the standard pattern, the characters of the X-ray film are pre-processed, and the characters are shifted vertically and horizontally, 5 sheets for each pattern, that is, 10 sheets in total. Figure
The eigenvalues obtained by adding SVD analysis determined by equation (7) to these standard patterns (in the upper left column in the figure), the correlation output values similarly determined in equation (7) (the upper right column in the figure),
Figure showing the coefficient (bottom left column and bottom right column in the figure) for the eigenvector u _i of the optimized SDF and the SVD of the conventional SDF
Shown in 13. The correlation output value of the conventional SDF is the pattern "0"
Is set to +1 and -1 is set to the pattern "1". As shown schematically in FIG. 14, the correlation output value for each pattern is obtained by projecting a pattern distributed in a multidimensional space onto an SDF formed so as to maximize the Fischer ratio, and a threshold for separating patterns. Can be analytically determined by examining the mean and variance of the distribution of correlation values on the projection axis. For that purpose, it is necessary to form the SDF with as many standard patterns as possible. Further, comparing the coefficients of the SDF with respect to the eigenvector u _i , the coefficient of the third vector or less (j ≧ 3) having a relatively small eigenvalue has the optimum SD in the conventional SDF.
It is larger than F. The reason why this component of the (+1, -1) output SDF is large is that the SDF is designed to output the same +1 or -1 correlation value for different patterns. The component having a small eigenvalue of SVD has low sensitivity of transmission to the correlation value, and is not important when performing discrimination. The increase in this component causes the noise component to expand. On the other hand, in the SDF formed so as to maximize the Fischer ratio, the fact that these components are kept small is a manifestation of optimization.

Next, an example of a specific method of forming an SDF using such a Fischer ratio as an evaluation standard will be described. If only two characters are to be identified, one SDF may be simply formed by using a standard pattern of two characters so that the Fischer ratio is maximized. When identifying 10 kinds of characters from 0 to 9 like numbers, it is necessary to specify a single number using a plurality of SDFs.

By the way, for example, the above 10 kinds of characters are classified into a character group of 2 classes by a classification procedure such as a so-called tree structure, and the classified characters are further classified into 2 classes, respectively. Into small classes,
There is a method of identifying these classifications by a composite identification function. Therefore, as a first method, an embodiment of a method for forming an SDF in such a tree structure will be described with reference to the schematic diagram of the identification tree structure shown in FIG. 15 and the flowchart of FIG. First, the SDF is formed so that the Fischer ratio is maximized for each of 10 combinations of two numbers, each of which is a number from "0" to "9" and a group of 9 numbers other than that. (Step 51).

Next, among the 10 SDFs formed, the SDF having the maximum Fischer ratio is selected (step 5
2). In the case of the present example, the SDF that identifies "7" and the other numeral groups has the maximum Fischer ratio, and the maximum value of the Fischer ratio is 35.1. Therefore, as the first SDF, the SDF that identifies “7” and the other number groups is selected.

Next, for one number identified by the selected composite discriminant function, that is, for the remaining number groups except "7", one number and eight other number groups other than the same as above. 9 SDFs are formed so that the Fischer ratio is maximized for each of the 9 combinations identified in the two classes, and the SDF that maximizes the Fischer ratio is selected (step 53). The second SDF selected at this time was an SDF that discriminates between “6” and other numeral groups, and the Fischer ratio was 17.5.

In the same manner, the SDF for discriminating one number from the remaining numbers and the remaining numbers are sequentially selected so as to maximize the Fisher ratio (step 54 ...
Step 70). The tenth SDF selected last is the SDF that identifies the two numbers. In this embodiment, "5"
It was an SDF that identified "3". The results were as shown in Figure 15. One number sequentially identified by SDF is "7" → "6" → "0" → "1" → "4" →
“2” → “8” → “9” → “5”, “3”, and each S
The Fischer ratio of DF is as shown.

By doing so, in each of the first to tenth discriminations, it is possible to obtain an SDF having the highest possible discrimination performance obtained by using the Fischer ratio as an evaluation criterion, whereby 0 to 9 are obtained. We now have 10 SDF groups that can best recognize the numbers. Next, an example of a method of recognizing character information using an SDF formed using the Fischer ratio as an evaluation standard will be described. In the present embodiment, an example in which the numeral of the ID number described in the character information section of the X-ray film is recognized by the 10 SDF groups formed as described above will be described with reference to the flowchart shown in FIG. To do.

First, after cutting out characters (numbers) from the character information section as described above, the correlation pattern x is obtained by performing a correlation operation with the first SDF for the inputted number pattern. The correlation value x is compared with a threshold value to identify whether the input numeral pattern belongs to "7" or other numeral groups (step 81). Then, when it is recognized as "7", it is stored that the number in the area of the corresponding character information portion is "7". At the same time, the first SDF
The probability of causing an error in the discrimination result by is determined based on the correlation value x. Specifically, as shown in FIG. 18, the probability density P1 in which the number corresponding to the correlation value x belongs to “7” by the first SDF and the number other than “7” have the highest correlation with “7”. Ratio P2 with probability density P2 belonging to high number "N"
Calculate by taking / P1. The value of P2 / P1 is 1 at the threshold value, and the closer to 1 the value is, the smaller the probability of correctly distinguishing two numbers, that is, it can be evaluated that the probability of an error in the identification result is high. , And the false recognition probability E is stored at the same time (step
82).

The input numeral pattern is the first SD.
When F is identified as belonging to a number group other than “7”, the value of P1 / P2 is used as the probability that an error has occurred in the identification result based on the correlation value x as described above, and the misidentification probability E
(Step 83). P1 in this case
Is a probability density belonging to “N”, and P2 is a probability density belonging to “7”.

Next, the input numeral pattern is the second SDF.
Correlation calculation is performed to identify which of "6" and the remaining numeral group belongs, and the identification result and the false recognition probability E are stored (steps 84 and 85). This operation is repeated until the input numeral pattern is recognized by a specific numeral (steps 85, 86 ... 105, 106). As mentioned above, each S
As the DF, one that maximizes the Fischer ratio is used in each identification, and thus high recognition performance is obtained.

In this way, the recognition of each input numeral pattern is completed, the ID number is determined by the combination, and after checking by the checksum etc., it is verified that there is an error by collating with the database. Sometimes it is necessary to re-identify as follows. For each recognized number, the number that is determined to have the highest probability of error based on the value of the misrecognition probability E stored at the time of identification is selected, and the value of E becomes the maximum for that number. The SDF identification result is found, the identification result is changed and identification is performed again, and another number is recognized. For example, when the number recognized as "2" in FIG. 15 is found to be incorrect (step 201), the values of the misrecognition probabilities E1 to E6 of the identification results of the first to sixth SDFs used for the recognition are compared. However, if the erroneous recognition probability E4 of the identification result by the fourth SDF is the maximum, the identification result is changed and recognized again as "1" (step 202). When there is a maximum value among E1 to E5, a specific number can be recognized by changing the identification result. If E6 is the maximum, the seventh S
It will be identified again from the DF.

In the above case, the number of "2" is changed to "1" and the collation is performed again.
Second, although the recognition error probability E is high, the identification result is changed and the identification is performed again. According to this method, the correction accuracy can be improved based on the error recognition probability E, and at the same time, the correction time can be shortened as much as possible.

In this embodiment, an SDF group having a tree structure for identifying one number and the remaining number group is used, but for example, 0-9 are the classes of five and five number groups first. To form a first SDF for identifying each of the five number groups, and a second and third SDF for identifying each of the five number group classes into two and three number group classes, and further identifying each class. You may make it form the SDF group of the tree structure which goes up.

According to such a tree structure recognition method, the number of times of recognition until recognition is significantly smaller than the number of characters, which has the advantage that the calculation time can be shortened. However, each character to be recognized and other characters There is also a method in which SDFs for identifying and are formed using the Fischer ratio as an evaluation criterion, and SDFs corresponding to the number of characters are formed corresponding to all characters. In this method, an SDF that maximizes the Fischer ratio may be formed using a standard pattern of a predetermined character image to be identified and a standard pattern of an image in which other characters are superimposed.

Character recognition using the SDF formed by such a method is as described in the embodiment of the preprocessing.

[0082]

As described above, according to the present invention
According to the character recognition method, the Fischer ratio is used as the evaluation standard.
For a tree structure for subclassifying from a large class to a small class or for distinguishing each character to be distinguished from other characters by using a formed composite function having a high classification function The identification function can be enhanced as much as possible.

Even if an error occurs in the recognition, a highly erroneous identification result is estimated, the result is changed, and the result is recognized again, so that the calculation time until the correct answer is obtained is minimized. Can be shortened .

[Brief description of drawings]

FIG. 1 is an overall configuration of a character recognition system according to the present invention,
Diagram showing functions

FIG. 2 is a diagram showing an example of a character information writing area in an X-ray film used in the above-mentioned embodiment.

FIG. 3 is a diagram showing a mode of an input state of the X-ray film.

FIG. 4 is a diagram showing a character information description portion and a result of projection processing in the embodiment.

FIG. 5 is a flowchart showing detection of a character information writing area according to the embodiment.

FIG. 6 is a flow chart showing the detection and correction of the inclination of a character.

FIG. 7 is a flowchart showing the first half of the process for detecting the direction and position of a character and correcting it normally.

FIG. 8 is a flowchart showing the latter half of the process following FIG.

FIG. 9 is a flowchart showing determination of a character cutout position and cutout processing.

FIG. 10 is a flowchart showing the same character recognition / collation processing.

FIG. 11 is a diagram showing a character identification function using SDF.

FIG. 12 is a diagram showing a projection state of a plurality of patterns onto an SDF that separates two classes.

FIG. 13 is a diagram comparing coefficients for the eigenvector u _i of the SDF optimized based on the Fischer ratio and the conventional SDF.

FIG. 14 is a diagram showing optimized SDF output values.

FIG. 15 is a schematic diagram of an SDF identification method for a tree structure.

FIG. 16 is a flowchart showing an embodiment of a method for forming an SDF for tree structure.

FIG. 17 is a flowchart showing an embodiment of a character recognition method using an SDF formed by the above method.

FIG. 18 is a diagram for explaining an erroneous recognition probability in the same embodiment.

[Explanation of symbols]

2 Processing computer F X-ray film

Front page continuation (72) Inventor Mitsui Matsui 1-26-2, Nishishinjuku, Shinjuku-ku, Tokyo Konica Corporation (56) Reference JP-A-58-166490 (JP, A) JP-A-63- 55677 (JP, A) JP-A-2-68684 (JP, A) JP-A-4-273590 (JP, A) JP-A-6-28525 (JP, A) Oyama Yamaaki Akira Yoshimoto Kensuke Yamaguchi, 1. ID recognition Research on Method Automatic Recognition of Film ID-Using Synthetic Discriminant Function and Data Matching Technology-, INNERVISION, Japan, Medical Science Co., Ltd., March 1993, Volume 8, No. 4, p. 34-37 (58) Fields investigated (Int.Cl. ⁷ , DB name) G06K 9/00-9/82

Claims (10)

(57) [Claims] 1. A type of character to be recognized is divided into two classes.
, And for each combination of the classification,
The synthetic discriminant function formed to maximize the Fischer ratio
Storing a group of numbers in a computer, the computer
The arithmetic processing is performed in the
-Composite knowledge corresponding to a combination of predetermined classifications with a large ratio
Select another function and then classify by the composite discriminant function 2
Reclassify two of the two classes again
And classify into the same class as above
Select the corresponding composite discriminant function and repeat the selection.
Characterized by returning to recognize all types of characters
Character recognition method. [Equation 1] 2. The group of composite discriminant functions is first 1
Forming a first composite discriminant function for discriminating one character from the remaining letters, and then discriminating one character from the rest of the characters and the rest of the characters excluding the character; characterized in that to form a, and is formed so as to be able to recognize all characters by repeating the same procedure the following
The character recognition method according to claim 1. 3. The group of composite discriminant functions is formed so as to maximize the Fischer ratio of each of the composite discriminant functions for discriminating between each character to be recognized and other characters .
The character recognition method according to claim 1, wherein: 4. A correlation calculation with a character to be recognized is performed by using each composite discriminant function , a class having a larger correlation value is selected, and classification and discrimination are performed from a large class to a small class. 請 characterized by recognizing a character
The character recognition method according to claim 1 or claim 2 . 5. The probability of causing an error in the discrimination result is calculated and stored based on the correlation value obtained by the correlation calculation with each synthetic discrimination function used until the character is recognized. The character recognition method according to claim 4 . When the 6. was misrecognized character is found, the probability of error in the stored identification result detects the highest synthetic discriminant function, again identified by changing the identification result character The character recognition method according to claim 5 , wherein the character recognition is performed again. 7. A method of performing a correlation operation between each synthetic discriminant function and a character to be recognized to obtain a correlation value respectively, and recognizing as a character corresponding to the largest correlation value among these. The character recognition method according to claim 3 . 8. The character recognition method according to claim 7, wherein when the character is found to be erroneously recognized, it is recognized again as a character corresponding to the next largest correlation value. 9. The character which is erroneously recognized is identified by checking the combination of characters obtained by recognizing each character with the recorded data of the correct combination of characters. The character recognition method according to any one of claims 4 to 8 . 10. A step of detecting a character area in which character information is recorded, a step of extracting character information from the detected character area, the extracted character information, and a group of the composite identification function.
Performs calculation of flops, and outputting a correlation value based on the respective operation result based on the output result of the correlation value, of the synthetic discriminant function
The step of selecting one from the group and specifying the character, the step of matching the specified character with the data in the database, and the step of outputting the composite discriminant function when there is a mismatch The character recognition method according to any one of claims 1 to 9, further comprising: converting the specified character into a different character based on the character.